Deliberation Memory: Preserving How Humans and Heterogeneous Agents Reach, Contest, and Revise Decisions

Long-running agentic work rarely advances as a sequence of stable facts. Humans and agents propose approaches, challenge assumptions, discover errors, revise plans, execute tools, and sometimes abandon earlier conclusions. Yet researchers commonly evaluate agent memory by whether it retrieves a relevant fact or improves a later answer. This framing can erase the trajectory needed to understand what a team believed, why it changed, and which evidence supported each change. We call the missing object deliberation memory: a persistent, provenance-preserving representation of observable acts through which humans and heterogeneous agents formulate, contest, revise, and act upon claims over time. We present an evidence-first architecture that separates append-only captured occurrences from rebuildable projections and retrieval indexes, distinguishes project identity from checkout provenance, and treats index freshness as an explicit correctness state. A local-first prototype has operated over 142,716 captured occurrences, 117,760 projected events, and 96,565 lexical retrieval units. Its development exposed two practical failure modes: silently splitting one project across repository checkouts, and repeatedly republishing an entire growing session after small updates. These experiences motivate a research agenda that measures memory quality through attribution, trajectory preservation, reconstructability, and honest disclosure of stale derived state, as well as retrieval relevance.